Adventures in AI Image Generation

Close up of my first attempt using a free AI image generator to create a hidden image puzzle.

Using AI to Create Images

A parent at Back to School Night for my Digital Art Class this year asked me if digital art was even a necessary skill anymore, given the rise of AI and its ability to generate images in seconds. It’s a valid question and one that made me realize that I need to explore AI image generation tools for myself and understand how they work and how they can be used in digital art or if they are poised to replace digital artists altogether.
As a digital art teacher, I’ve done some experimentation with the generative tools in Photoshop, but diving deeper into AI-driven image generation felt like the next step. I decided to start my exploration with a hidden image puzzle. I've always liked these types of puzzles and thought I could create a puzzle that my art club students would enjoy if we finished our project early. 
I had the idea of hiding art materials into a crowded art gallery scene. I was imagining a 19th Century Salon style exhibition gallery with paintings hung floor to ceiling and a crowd of viewers in the foreground among pedestals with sculptures. My plan was to hide a variety of art making tools in the scene and it would be a fun way to review the names of different tools as well. 
This hidden image puzzle project was a way for me to dip my toes into AI-driven art and see what the free version of this technology is capable of at this point. Through this process, I discovered that digital skills are still very much needed, even with AI in the picture.

Rowlandson, Thomas. Royal Academy Exhibition Room, Somerset House. Copyright Royal Academy of Arts

How to Write Prompts for AI Image Generation

The first step in creating a hidden image puzzle using AI is outlining the concept. In this case, I wanted to design a puzzle set inside an art gallery, where viewers could search for hidden art tools—paint brushes, a palette, a pencil, a box of pastels, a printer's brayer, a dip pen, an inkwell and other art supplies. My vision included a gallery scene with these tools subtly blended into the paintings, sculptures, and the gallery’s architectural details.
I logged into ChatGPT and selected the "web image" generation option. I typed in the prompt "generate a hidden image puzzle with art supplies hidden in a salon-style gallery scene. Please include an answer key along the bottom of the image with images of the hidden objects. Hidden objects should include: a pencil, a paintbrush, an inkwell, a dip pen, a set of artist pastels, a sketchbook, a printmaking brayer, a set of paints in a box  and a palette." 
The result of this first experiment was almost nightmarish in its distortions, reminding me of those AI generated scenes of semi-familiar objects that cannot be identified created to simulate what it feels like to suffer a stroke. 
The scene, recognizable as an art gallery, included lumpy forms that somewhat resembled sculptures on pedestals as wells as a number of more ambiguous forms in the foreground. The "hidden image" key was just as muddled. It included 17 strange objects with loose visual ties to 10 specific art materials in the prompt and none of these objects are actually replicated in the scene, defeating the point of it being a hidden image puzzle. 

Full AI image generated by my first prompt to create a hidden image puzzle in a salon style art gallery.

Refining the AI Image Generation Prompt

I wasn't completely disappointed with this first iteration but it definitely missed the mark which meant I needed to focus on crafting a better prompt for the free ChatGPT AI Image Generator.  
I decided to look carefully at what the AI was doing well to start to make sure I kept the key aspects of my prompt that were working. The scene communicates the salon-style I was going for and some of the structure of the image is correct including the art on the walls, pedestals with sculptures, and the answer key along the bottom. 
Then I thought about what was not working in the image: the objects are not recognizable, the objects in the answer key don't match the objects in the scene. Also, I wanted to clear the gallery floor from extra objects that didn't relate to the puzzle. 
I was feeling much more confident heading into the the next prompt. I decided to give the AI a particular type of art to focus on rather than just say "art" I used the keywords "landscape and figure sculpture" and I included more detail about what is meant by "hidden images." I also  
My new prompt: 
"Generate a hidden image puzzle with art supplies hidden in a salon-style gallery scene. The gallery should include landscape painting and figure sculptures on pedestals. Please include an answer key along the bottom of the image with images of these objects: a pencil, a paintbrush, an inkwell, a dip pen, a set of artist pastels, a sketchbook, a printmaking brayer, a set of paints in a box  and a palette. These objects must be hidden in the scene and be identical to answer key. Please do not include extra objects on the floor if they are not a hidden image."  ​

My second attempt to use free AI Image Generation to create a hidden image puzzle.

As you can see in the image above, which was the result of this second attempt, there are some improvements: The artwork is less distorted and the gallery floor is clear of extra objects. The answer key has less objects but still not the 10 objects I listed. 
But the main issue remains, the objects are not hidden with in the scene and the objects themselves are not recognizable. 
I will spare you the details of the lengthy prompts I created with subtle changes to my prompt...I ended up adding greater description to the objects themselves to make them more recognizable but I found as the AI began to grasp these details it let go of others, including the structure of the answer key and the concept of a hidden image. 
In this iteration below I used several sentences to describe what is meant by portrait and figure sculpture and landscape art. I find it interesting that the AI generated images reminiscent to artworks of art history like Van Gogh's Starry Night and the Venus de Milo. 
I also carefully described the 10 art supplies I wanted in the scene but it seems as if the AI focused solely on the painting objects. Most importantly the objects hidden in the scene are not identical to the objects in the foreground. 

My third attempt at using free IA Image Generation to create a hidden image puzzle.

I made subtle edits to my prompt and generated images with similar issues. It seemed like whenever I was able to resolve one problem the AI created another issue. Although the images became more refined and less amorphous and nightmarish the goal of creating a hidden image puzzle was elusive. Below was another iteration where the hidden images appear more clearly in a "key" but are not identical to the images that are "hidden." 

My fourth attempt at using free AI image generation to create a hidden object puzzle.

Last Attempt at Using AI to Generate a Hidden Image Puzzle

At this point after creating several iterations of the prompt including those with specific instructions and descriptions I decided to try once more by attempting to make the scene more crowded. I figured maybe the AI was having an issue hiding the objects because there needed to be more opportunities to hide the objects. I decided to explicitly ask for a salon-style exhibition with people to see if the objects could be hidden. I also included the instructions to number the hidden objects hoping this would limit them to the 10 I described. I also let go of the lengthy descriptions to try to focus the efforts toward actually hiding identical objects in the scene. 
Here's the final prompt: "Generate a hidden image puzzle with art supplies hidden in a salon-style gallery scene with art on the walls and on pedestals and viewers walking around the scene. The gallery should include landscape paintings and figure sculptures on pedestals. Please include a numbered answer key along the bottom of the image with these 10 images: a pencil, a paintbrush, an inkwell, a dip pen, a set of artist pastels, a sketchbook, a printmaking brayer, a set of paints in a box and a palette. Identical copies of these 10 images must be hidden in the scene and be in every way identical to images in the answer key.​"

My final attempt in using free AI image generation to create a hidden object puzzle.

I'll let you be the judge of the results. I'm not sure this worked any better than the previous attempts. 
Although my efforts to use free AI (DALL-E) image generation to create a hidden object puzzle failed. I did get to see how these powerful tools work and also get temperature check on how accurate they actually are (at least for the free version.) 
My only question now is how is this free CHAT GPT DALL-E version different from some of the paid options out there...but that will have to be continued because I'm not convinced I want to spend money on this service at this point. 

One More AI Image Generation Experiment

Before putting this AI image generation experiment on the back-burner for a while I decided to return to my original intention. To see if the free AI image generation tools were going to replace the need for Digital Art classes and put me out of the job as a Digital Art teacher...
For the last experiment I decided to craft a prompt to replicate my example image for one of my favorite Adobe Photoshop lessons for beginners: Digital Surreal Collage. 
This lesson introduces students to the genre and movement of Surrealism, they learn the surrealist strategies and some of the key image editing tools/techniques in Adobe Photoshop including the selection tools, transform tools, layer effects, and adjustment layers. It is important for me to teach my students how to use found images responsibly and how to discern what is fair use online. Because of this I direct my students to use copyright free image sites like Pexels.com and to use image free to use archives like the US Library of Congress Free to Use and Reuse Image Gallery.  

My Digital Surreal Collage example created in Adobe Photoshop

Digital Surreal Collage Without AI

I created my example image for the Digital Surreal Collage lesson using both Pexels.com and the Library of Congress images, selecting a vintage photo of a female tennis player as my inspiration created a scene where I employed dislocation placing the tennis player in the sky surrounded by clouds (a nod to my favorite surrealist Rene Magritte) and I used change of scale to make the ball oversized symbolizing the sun.
I used several of the key tools and techniques in Adobe Photoshop to create the image including layer effects (drop shadow), adjustment layers (to create color adjustments), clone stamp and transform tools. It was a relatively easy image to create and only took 10 min to create but I wanted to see if the free DALL-E AI generator could create a similar image instantly... 

AI Generated Surrealism

I did my best to craft a prompt using what I had learned from my failed hidden image puzzle experiment. This is what I came up with:
"Generate an image of a female tennis player from the early 1900s in black and white playing tennis in the clouds with a giant tennis ball. Put a drop shadow behind the tennis player and make the clouds pale pink, purple and sky blue."
My first impression was that this image came out much more successfully the more complicated hidden object puzzles. But clearly there are some issues: the entire image came out black and white so I think I would make sure to specifically ask for only the tennis player to be in black and white because the cloud colors specified here were completely ignored probably because the black and white request took precedence. 
Also if you look at the image more closely you will notice that the cloud shape is replicated several times at different scales in the scene making it look fake. Also the image is missing the drop shadow behind the tennis player. 

Is AI Good at Creating Images?

Having completed this mini-experiment (I still feel like I've only scratched the surface) into AI image generation, I have more insights into what the free image generator can and cannot do. 
Do I think the free AI image generator I used (DALL-E) is good at creating images? No. I don't, not at this time. If students used this tool to create their work it would be obvious. The irregularities and strangeness pervades the image and is hard to overlook. 
The images I was able to create are intriguing and at times frightening but they do not have the clarity, complexity and intentionality of digital art created by humans using applications like Adobe Photoshop. So for now I think my job is safe and I still think teaching digital art is valuable for students.